White pulp segmentation algorithm for mouse spleen cryo-imaging data using U-net

Abstract

Spleen is one of the important organs for studying graft-versus-host disease development using cryo-imaging modality. Up to this point, there is no algorithm developed to segment spleen tissues in the cryo-imaging data. In this study, we developed a new, automatic segmentation algorithm for spleen tissues in the cryo-images for the first time. The algorithm consisted of (1) pre-processing, (2) predicting, and (3) post-processing. The predicting model consisted of two U-Nets to separate spleen from background and white pulp from red pulp. The results generated by the algorithm were compared against the ground truth generated by cryo-imaging experts. The Dice similarity coefficients were about 86% for red pulp class and 87% for white pulp class. We also used the algorithm to measure spleen volumes and performed the T-cells proliferation assays on the results. We found that there was a significant difference between ratios of white pulp volume to spleen volume between allogeneic spleen and syngeneic spleen. The results were consistent with the previous biological reports. In conclusion, the algorithm accurately and objectively segmented the spleen tissues with similar quality to the experts. By incorporating the algorithm into the T-cell proliferation assay workflow, it should greatly increase throughput of the process as compared to manual segmentation by humans. We expect that this algorithm will help accelerate the research for further understanding of graft-versus-host disease and development of the cure.